Understanding the Different Claudes

Understanding the Different Claudes

What is the difference between Claude at claude.ai, the Claude desktop app, and Claude Code? A way to tell them apart: ask where the work ends. A chat ends at an answer, work handed over with a connected folder ends at finished files, and Code work ends when the project runs or passes its checks. Then check which files and connectors it can reach, and which computer actually runs the commands. Part of the series Claude with Professor Bear. Liam, in for Bear, narrates.

Chapters 0:00 Opening 0:08 The idea 0:18 Terms in this lecture 0:31 I · Names and Surfaces 1:38 II · Files and Context 2:48 III · Where Work Runs 4:19 IV · Choose for the Job 7:27 Recap 7:37 Your turn

Your turn. Paste this into Claude: Help me choose a Claude setup for one task. Ask what I want to produce, where the work ends, where the files live, and where the required tools are installed. Recommend an interface and execution environment, then give me a small test and a way to check the saved result. Then check two things yourself: does the host you chose have the files and tools? Does the result exist where you expected it?

Product interfaces vary by account, configuration, and rollout. Research and interface captures checked October 12, 2026. The remote-terminal commands shown are an illustration and were not executed on a cluster. The two-account workflow near the end is Bear's own plan, not a tested configuration.

Sources

• Claude, Cowork, and chat are one Claude: https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude
• Chat search and memory: https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context
• Claude Code overview: https://code.claude.com/docs/en/overview
• Claude Code desktop: https://code.claude.com/docs/en/desktop
• Claude Code memory: https://code.claude.com/docs/en/memory
• Claude Code Remote Control: https://code.claude.com/docs/en/remote-control
• Connect Claude Code to tools via MCP: https://code.claude.com/docs/en/mcp
• Get started with connectors: https://claude.com/docs/connectors/overview

Made with the brutalist.art toolkit: Manim, Remotion and Kokoro, free and local. Narration is synthetic (Liam, in for Bear). This is an independent educational film, not an official Anthropic publication.

What this video covers

• 0:00 — Opening
• 0:08 — The idea
• 0:18 — Terms in this lecture
• 0:31 — I · Names and Surfaces
• 1:38 — II · Files and Context
• 2:48 — III · Where Work Runs
• 4:19 — IV · Choose for the Job
• 7:27 — Recap
• 7:37 — Your turn

Transcript

A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.

Opening

[0:00] Claude with Professor Bear, understanding the different Clauds. Choose the right interface, connect the right files, and know where the work runs.

The idea

[0:08] Hello, this is Liam in for Bear. These names overlap. Start with the task, then ask which interface you need, which files it can reach, and which computer will do the work. An interface is where you interact, an

Terms in this lecture

[0:19] Environment is where the tools run. A terminal is a text window for commands. Context is the information available to the current task. A connector is a link that lets Claude use another service.

I · Names and Surfaces

[0:31] First, separate the names. A website, an installed app, and a coding agent are different kinds of things, even when all three carry the Claude name. Claude.ai is the website. Open it for conversation, research, writing, and working with information you provide. Calling it Claude AI does not make it a separate model from desktop Claude. Claude desktop is the installed application. It gives you a desktop interface to Claude, and where available, connections to your local work. The application can also contain a separate code mode. Power capabilities are being brought into the main Claude experience. That explains the changing labels. Your account may still show an older arrangement, so teach the capability and inspect the interface you actually have. Claude code is an agent that can inspect a project, change files, and run commands. You can use it through a terminal or a desktop interface, with other supported interfaces available, too. The useful distinction is the working setup. Web Claude centers the conversation. Desktop adds local connections. Code centers work in a project and its tools. These capabilities overlap. The names are not separate intelligence levels.

II · Files and Context

[1:38] Next, ask what information the task can actually use. A remembered preference, an uploaded document, and a connected project folder are three different sources of context. With web Claude, upload material or use an available connector. A file in a cloud workspace is not automatically the live file on your laptop. Check where an edited result will be saved. In desktop Claude, you can connect a folder when the feature is available and grant access to its files. That makes local reading and writing possible. The folder connection is a permission boundary. Connecting a folder does not by itself tell you where commands run. If the job needs your installed renderer or Python environment, verify the execution environment instead of inferring it from the visible files. Chat history and enabled chat memory can follow your account across web, desktop, and mobile. Code has its own project instructions and auto memory. Do not assume every conversation or remember detail automatically crosses those boundaries. Both chat and code can help produce code and documents. The important question is what happens next, an artifact to review, a file to download, or a change written into a working project. Now identify the

III · Where Work Runs

[2:48] Machine doing the work. Local, cloud, and remote are execution choices. They determine which files, installed tools, and credentials the task can actually use. The desktop code environment selector makes this choice visible. Local uses this computer. Cloud uses a hosted environment. SSH connects to a remote machine. Desktop code is therefore another route to remote work. For a render on your Mac, choose a setup that runs tools on that Mac. The renderer, Python environment, files, and credentials must be available there. Local execution still normally contacts a remote Claude model. For a Linux box or an HPC cluster, a terminal is often the practical route. Connect to the remote machine, enter the project, and start code there once access, installation, and network

Requirements are satisfied. In that workflow, code reads and changes files on the remote host. Your laptop supplies the interface. You do not need a graphical desktop on the Linux machine to use its terminal installation. Code can also run non-interactively. Its own help describes the print option for that purpose. Automation still needs permissions, inputs, and a process that can finish. A command line interface does not guarantee uninterrupted execution. Cloud work can continue while your laptop is off. A local task needs your computer to stay on, and it uses that computer's files and tools. A terminal task over SSH needs its remote host and process to keep running. Remote control lets another device view a local code session without moving that work.

IV · Choose for the Job

[4:19] Put the distinctions to work. Start with the result you need, locate its files and tools, and then choose the interface that gives you the right access. Here's a cleaner way to tell them apart. Ask where the work ends. A chat ends when it gives you an answer. Work handed over with a connected folder ends when the finished files exist. Code work ends when the project actually runs or passes its checks. So, ask, hand over, build. Three signals tell you that you are in the wrong place. If you keep copying answers out of a chat into documents, hand that assembly work over with a folder. If your thought is, "I wish a tool existed for this." that is a build, so use code. And if the job needs software installed on your machine, it needs code running there. For researching a topic

Or shaping an explanation, begin with Claude chat in the browser or desktop. Use the sources and available account context, then check the important claims against evidence. Now, take two jobs on the same folder. The first is writing a script into it. That work ends when the files exist and read well, so desktop Claude with the folder connected is usually enough. The second job is a repair. A quality check flags a problem. You change one setting, rebuild one piece, and run the check again until it passes. Running commands and rechecking is the heart of that loop, and it is what code is for. Connectors are a separate question. A connector is a link that lets Claude use another service, such as a design tool, a mailbox, or an image generator. Many are built on an open standard called MCP. Desktop Claude and

Code can both use connectors, but a connector belongs to the account and the setup that connected it. With two accounts, check which one holds the connection. A connector tells you a Claude can reach, not where its commands run. For rendering the film, use local code or a terminal on the machine with the rendering pipeline. A well-written script is an input. The installed software and the generated media are separate things to verify. For the Linux or HPC workflow, use terminal code on the remote machine or desktop code with SSH. The deciding factor is where the project and tools live, not whether you prefer a window or a prompt. When planning and execution happen on different machines, save a clear handoff in the project and transfer the intended version. Name the inputs, commands, outputs, and checks so the executor does not depend on

Remembered conversation. Bear has two Claude accounts, one through Northeastern and a personal Max account. His workflow can use Northeastern access for terminal production and Max for interactive desktop work. Confirm the active account before starting each stage. His book loop and film loop scripts show the production pattern. A shell loop starts fresh Claude code jobs and saves progress on disk. Headless execution can produce rough drafts without keeping an interactive conversation open. Then open the saved project in desktop for hands-on refinement. Inspect the rough cut, revise the script, and request changes. Carry over the source files, assets, and build notes. Switching accounts does not transfer the conversation or combine allowances. Before starting, check the selected environment, connected folders, and available tools. Browser or app control depends on configured integrations and permissions. Then make a small reversible test and inspect the actual saved result.

Recap

[7:28] Let's recap with Claude. Ask start with a chat. Hand over desktop Claude with a folder. Build code on the machine with the tools. Then check where it runs and what it can reach.

Your turn

[7:38] Help me choose a Claude setup for one task. Ask what I want to produce, where the work ends, where the files live, and where the required tools are installed. Recommend an interface and execution environment, then give me a small test and a way to check the saved result. Understanding the different clauses at Nick Bear Brown.